Matei Ripeanu
Tuesday, April 19, 2005
1127/1131 Kemper Hall
3:10-4:00 p.m.
Abstract:
Understanding how to build self-organizing systems is paramount for coping efficiently with the increased scale and complexity of the computing infrastructures we envision. Existing deployments of Grid infrastructure provide the real world requirements for such an exploration and can be used to evaluate solutions in realistic settings.
Group communication primitives have broad utility as building blocks for such infrastructures. However, creating and maintaining the distributed structures that support these primitives is made difficult by transient networks and large scale. In this talk I present a solution that supports group communication primitives on large and dynamic sets of Internet-connected hosts. This solution uses an unstructured overlay and a novel approach to extract source-specific, data distribution trees. Its strength derives from the following two properties: First, it decouples overlay construction and maintenance mechanisms and tree-extraction mechanisms, allowing for separate component optimization. Second, it relies on soft-state and passive data collection to adapt to the dynamics of the physical network, resulting in a self-organizing, low-complexity system.
I will present experimental and analytical evaluations that demonstrate low communication overhead, efficient network usage (low delay penalty and network stress), and ability to adapt quickly to network changes and recover from node failures. If time permits, I will present a replica-location service built on top of this group communication layer that proves the practicality of this approach
Bio
Matei Ripeanu is a Ph.D. candidate in Computer Science at The University
of Chicago. Matei is broadly interested in large-scale distributed systems.
During his time at The University of Chicago, Matei has published articles
on large Grid and peer-to-peer system characterization, on techniques
to exploit the emerging characteristics of these systems, and on supporting
scientific applications to run on these platforms. Matei is a recipient
of the Gordon Bell Award (2001) for "significant achievements in
applying high-performance computers to scientific and engineering problems."
For more information please visit: http://www.cs.uchicago.edu/~matei